about adding a measurement to telemetry

[Yanbo Xing]

Hello,

 

I want to add a measurement about scrolling performance of android chrome, the details are as follows:

1.       Do a fling scroll when the browser loads a heavy web site for the first time.

2.       How many tiles are successfully rendering during fling scroll ? How many tiles are not ready and with blank content during fling scroll ? And calculate the tile percentage ((tiles rendered successfully)/(total  number of tiles)) during fling scroll.

 

Can we add this case? And how to get started ?

 

Thanks

Yanbo

Marvell (Shang hai) LTD

[Yanbo Xing]

[Nat Duca]

Wei is working on something related, maybe you can use her work.

[Weiliang Chen]

I have done a Histogram data tracking for number of checkerboard or low res tiles when in compositor thread interaction (fling scroll, animation, pinch zoom, etc). Details of how to count tiles can be found here: https://codereview.chromium.org/364063005/diff/140001/cc/trees/layer_tree_host_impl.cc

On the other hand Chris has done telemetry test about checkerboard/low res area (here I believe: https://codereview.chromium.org/1033733010)

I cc'ed him too.

Thanks,

Wei

[Chris Blume]

That is correct. There is already a test in place to determine which tiles were checkerboarded (no content). And there is a separate test for which tiles were approximated (both no content and low-res content). If you subtract one from the other you can determine how many tiles were just low-res. Either way, it saves you having to write a telemetry test to determine the percentage.

There is also a page set to test fling behavior, which Jared wrote. I've included him to the email as well.

So the part which isn't covered is "for the first time." I believe you are talking about finding behavior of flinging while content is still downloading.

I don't currently know of a reliable way to test that. I know we save test pages and test against the saved copy instead of a downloaded copy. That makes sure that if a page changes we aren't testing against new content and thus ruining our test result history. I do not know if the saved content is loaded slowly to emulate being downloaded over a network.

Chris Blume | Software Engineer | cbl...@google.com | +1-614-929-9221

[Jared Duke]

There is also a page set to test fling behavior, which Jared wrote. I've included him to the email as well.

Right, for fling gestures we currently have smoothness.fling.simple_mobile_sites. For that test suite, we're more interested in raw fling animation performance than responsiveness during page load, so the sites are intentionally somewhat simple and we explicitly wait until the page is considered "loaded" before initiating the gesture.

Measuring gesture performance *during* page load is hard for a number of reasons, but the biggest problem is noise. There's often simply too much variability between runs for the data to be actionable on a continual basis. Historically we've grouped sites that are noisy and/or exhibit stressful loading symptoms long after the load signal in our smoothness.pathological_mobile_sites suite. I wouldn't be opposed to adding another such suite for fling, though I'm curious if you're interested in the actual fling animation throughput or just the initial touch response latency?

[Yanbo Xing]

Hello Jared,

 

       We are interested in the fling scrolling performance after the page loaded(not during page load).  “for the first time”  means that browser gets the tile content through the drawing texture(not form the cache).

       I will check the fling gestures and telemetry test about checkerboard/low res area.

 

      Thanks for your quick reply!

 

Best Wishes!

Yanbo